Skip to content

Systems and records

How should an agent resolve conflicting refund rules? A retail task packet

By SourceX Editorial · Reviewed by Noah Loul ·

9 sections

Short answer

An agent should rank refund rules before applying them: mandatory consumer law first, then the policy version disclosed at purchase, then item flags such as final sale, then promotion terms, then approved goodwill. A retail refund policy conflict dataset tests this by packaging each disputed refund with its order lines, rules in force, jurisdiction facts, the decision and later refund or return evidence.

Agent run

Retail refund policy

This page covers how should an agent resolve conflicting refund rules? A retail task packet.Illustrative

Key takeaways

  • Rank refund rules before applying them, and treat a tie on the same tier as a handoff rather than a guess.
  • Store the policy version, item flags and promotion terms as they stood at purchase; today's settings leak the wrong rule into old tasks.
  • Statutes such as the EU withdrawal right and California's display duty can override a store window that looks clear.
  • Grade the refund and line statuses the agent changed, then the rule it cited; a polite closing message is not an outcome.
  • Pair each conflict task with a counterexample that differs in one fact, so a model has to read the records.

How should an agent resolve conflicting refund eligibility rules?

It should rank the rules before it applies any of them: a refund right the law gives the customer comes first, then the policy version that was in force and shown to the customer at purchase, then item-level flags such as final sale or a shorter window, then promotion terms, and only then discretionary goodwill that needs an approver. The agent decides from the records, names the rule that controlled, and hands off to a person when two rules sit on the same tier.

That answer is easy to state and hard to test. Many refund transcripts end with a polite message, and a model can learn to produce the message without ever checking which rule applied. A retail refund policy conflict dataset fixes this by packaging each disputed refund as a task with its starting state, every rule that could apply, the linked records that decide between them, a recorded decision and evidence of what actually happened afterwards.

This page gives a field-level design for that packet, a precedence ladder you can adapt, an illustrative worked case with a counterexample, and the checks that keep the packet from teaching the wrong lesson. It is written for retail and ecommerce operations and data leads who hold years of order, return and refund history and want to know what a usable conflict task looks like.

Where do refund rules actually collide in a store's records?

Conflicts appear wherever two sources of truth describe the same purchase differently: the store-wide policy and a product override, the posted terms and a statute, the order record and a promotion's fine print. The table lists the collisions worth sampling, with the record that exposes each one.

Platform behavior matters here because it is part of the rule set the agent inherits. Shopify's help center, for example, says that when an item falls under several return window overrides the shortest one applies, and that an item marked final sale stays final sale even if an override also covers it. An agent trained on a store that runs those settings must reproduce that logic, not a generic one.

Where do refund rules actually collide in a store's records?
CollisionHow it shows up (illustrative)Record that exposes itWhat usually decides it
Store window vs legal withdrawal rightA 10-day store window for an EU online buyer who asks on day 12Ship-to country, delivery scan date, request timestampThe statutory period where it applies; the store window cannot shorten it
Unposted no-refund policyA California store refuses a refund but never displayed the policySignage or checkout evidence for that policy versionWhether the display duty was met for that sale
Final sale vs return window overrideA clearance item inside a collection with a 60-day windowProduct flags and collection rules as of purchase datePlatform setting: final sale wins in Shopify return rules
Two overlapping overridesA product in two collections with 30-day and 14-day windowsCollection membership historyPlatform setting: the shortest window applies in Shopify
Delayed shipment vs no-cancellation clauseAn order not shipped by the promised date, customer asks to cancelPromised ship date, shipment events, delay noticesThe FTC mail order rule's cancel-and-refund option
Policy changed between purchase and requestWindow cut from 30 to 14 days a week after the orderPolicy version log with effective datesUsually the version accepted at purchase, per company rule
Promotion terms vs standard returnsA gift-with-purchase must come back with the main itemPromotion record linked to the order linesPromotion terms if they were disclosed at checkout
Refund method restrictionCustomer wants a card refund for a gift card paymentOriginal tender and refund transactionTender rules in policy, unless a statute overrides them

Which rule wins: a precedence ladder an agent can follow

Write the precedence down before you label a single task, because the label depends on it. The ladder below is an editorial starting point, not legal advice; the company's counsel should confirm each tier for the jurisdictions it sells into, and the confirmed version becomes part of the packet's documentation.

Three statutes show why law sits on top. The EU Consumer Rights Directive gives consumers 14 days to withdraw from a distance contract and requires the trader to reimburse payments, including delivery costs, no later than 14 days after being told of the withdrawal, though for goods the trader may withhold reimbursement until the goods or proof of return arrive. California Civil Code section 1723 requires a retailer whose policy denies full refunds or equal exchanges for at least seven days after purchase to display that policy, and makes a retailer that fails to do so liable for the purchase price when goods are returned within 30 days. The FTC's guide to the Mail, Internet, or Telephone Order Merchandise Rule says a cancelled order must generally be refunded within seven working days, or within one billing cycle when the seller itself extended the credit.

Each of these can override a store policy that looks clear on its own. That is exactly the kind of interaction a model will miss if the training set holds only the policy text and the final reply.

  • Tier 1, mandatory law for the customer's location and channel: withdrawal rights, display duties, delayed-shipment cancellation rights.
  • Tier 2, the policy version in force and disclosed at the time of purchase, with its effective date and where it was shown.
  • Tier 3, item-level settings: final sale flags, product or collection window overrides, bundle rules.
  • Tier 4, promotion terms attached to the order lines, if disclosed at checkout.
  • Tier 5, discretionary goodwill or exceptions, valid only with the approval the policy names.
  • Tie rule: when two rules on the same tier disagree, or the record cannot show which applied, the correct action is a documented handoff, not a guess.

What should the policy conflict retail task packet include?

Each task is one refund request that touched at least two rules, with everything a careful reviewer would need to decide it and confirm the outcome. The field groups below can be filled from an order management system, the helpdesk, the payments ledger and the policy archive; a team can use them as the packet schema as is.

At a glance

What should the policy conflict retail task packet include

Keep times in one stated time zone with offsets preserved. A refund window that closes at midnight in the store's admin time zone can open or close a day early for a customer elsewhere, and a packet that drops the offset makes some tasks unanswerable.

What should the policy conflict retail task packet include?
Field groupFieldsSource recordWhy it is needed
Task headerTask id, conflict type code, packet version, label status (real or illustrative)Packet build logLets reviewers sample by conflict type and trace each label
Order and item linksOrder id, line ids, product and variant ids, bundle membership, tender per lineOrder management or ecommerce platformRefund eligibility is decided per line, not per order
TimelinePurchase, ship, per-item delivery, request and decision timestamps with offsetsOrder events, carrier scans, helpdeskWindows start from delivery or purchase depending on the rule
Rules in forcePolicy version id, effective dates, disclosure evidence, override and final sale flags as of purchasePolicy archive, product settings historyThe agent must apply the rule that applied then, not today's
Jurisdiction factsShip-to country and state, sales channel, business or consumer buyerOrder address, account typeDecides whether a statutory tier applies at all
Promotion linksPromotion ids and terms attached to each line, free items, thresholdsPromotions engine, order discount linesRemoving a qualifying item can change the refund amount
ConversationCustomer messages and agent replies, de-identified, with timestampsHelpdesk such as Zendesk or GorgiasShows what was asked and what was promised
Decision and citationAction taken, rule tier and clause cited, approver if any, handoff flagAgent notes, approval logThe label: which rule controlled and why
Outcome evidenceRefund transaction, return receipt scan, chargeback or complaint within a set periodPayments ledger, warehouse, disputesConfirms the decision held, beyond the status field

Worked case (illustrative): a promotional sweater returned on day 12

The case below is illustrative and does not describe a real customer or store. Durations are stated in days and no money amounts are used. It shows the packet answer format: starting state, linked evidence, decision and verified outcome.

  • Starting state: one order with two lines, a sweater and a free scarf from a buy-one-get-one promotion; both delivered on day 0 to an address in an EU member state; customer asks to return the sweater on day 12.
  • Rules in force: store policy version 7 with a 10-day window from delivery, disclosed at checkout; the sweater is not final sale; the promotion terms say the free item must be returned with the qualifying item.
  • Apparent conflict: the store window has closed, but the customer is inside a 14-day statutory withdrawal period for a distance sale.
  • Linked evidence: carrier delivery scan for both lines, the request timestamp, the ship-to country, the promotion record on the order lines and the policy version log.
  • Decision: accept the return under tier 1, require the scarf back under tier 4, refund to the original tender, cite the statutory tier and the promotion clause.
  • Outcome check: return receipt scan for both items, refund transaction posted within the statutory period after notice, and no complaint or chargeback in the following 60 days (an illustrative follow-up window).
  • What a weak agent does: refuses on the 10-day window and writes a courteous message, which a transcript-only grader might score as resolved.
  • Counterexample: in a sibling task the same customer buys the same sweater from the store's U.S. site, shipped to a state with no relevant display rule, and the 10-day policy was shown at checkout. Tier 1 does not apply, tier 3 does not apply, and tier 2 controls, so a refund refusal with an offer of an exchange is the correct action. Pairing the two tasks forces a model to read the ship-to and delivery fields instead of memorizing that late sweater returns are refused.

How do you grade the agent on the refund, not the reply?

Grade the final state of the systems the agent changed, then grade the citation. The tau-bench retail benchmark is a useful public reference: its environment code computes a hash of the database after the agent's run and compares it with the state produced by replaying the correct actions, and a mismatch sets the reward to zero. Its retail policy also encodes conflict-relevant constraints, such as allowing returns only for delivered orders and sending refunds to the original payment method or an existing gift card.

At a glance

How do you grade the agent on the refund, not the reply

The same benchmark reports pass^k, the chance that an agent succeeds on all of k repeated trials of the same task, alongside single-run success. Repeated trials matter for refund conflicts because an agent that applies the right tier only some of the time will still look competent in a one-shot test.

For records from a real store, the end state is the refund or return transaction and the order's line statuses, not the helpdesk ticket status. A ticket marked solved after a refusal is a completed conversation, not evidence that the correct rule was applied.

  • State check: line statuses, return authorization and refund transaction match the labeled decision.
  • Amount check: refunded lines and promotion adjustments match the decision, expressed as which lines were refunded.
  • Citation check: the tier and clause the agent names match the label.
  • Handoff check: tasks labeled as ties are escalated, and the agent does not invent a rule to break the tie.
  • Reliability check: run each task several times and report pass^k next to single-run success.

Which records quietly break a refund conflict packet?

Leakage and missing history can break a packet even when every label is right. The patterns below are the ones to test for before a packet is used for training or evaluation.

Treat a missing policy version as a reason to drop the task, not to fill in today's policy. A packet built on reconstructed rules teaches a model to apply a rule the customer never saw.

  • Policy drift: product flags and collection rules exported as they are today instead of as of the purchase date.
  • Future leakage: the refund transaction or a later complaint is visible in the starting state given to the agent.
  • Status as truth: a ticket status or a refund issued by a supervisor treated as proof that the first answer was correct.
  • Silent goodwill: discretionary refunds recorded without an approver, which make tier 5 look like the default.
  • Unresolved ties labeled as settled: if staff disagreed or the record is thin, keep the disagreement and label the task as a handoff.
  • Skewed sampling: routine in-window returns dominate history; sample by conflict type so rare collisions appear often enough to measure.

What has to be settled before refund histories leave the company?

Refund cases carry names, addresses, order numbers, payment details and free-text complaints, and none of that is needed to decide which rule applied. Replace customer and order identifiers with consistent tokens, reduce addresses to the jurisdiction fields the precedence ladder uses, drop card and bank details entirely, and review free text for health, family or other sensitive details customers mention when they explain a return.

Privacy work and licensing rights are separate questions. De-identification reduces what the records reveal; it does not by itself give the company the right to license them. Policy documents may be the company's own, but helpdesk exports, marketplace orders and platform data can be subject to customer terms, marketplace agreements or vendor contracts. Counsel reviews the specific facts, and the precedence ladder itself should be confirmed by counsel before any label depends on it.

What a retailer brings to a refund conflict license

A retailer with several years of linked orders, policy versions, return scans and refund transactions holds the evidence these tasks need. On the licensing side, SourceX takes a retailer from qualification and rights review through contracting, preparation, delivery and payment.

No data is needed to check fit; a short company assessment comes first. A company that proceeds approves scope and terms and signs before anything is delivered, and SourceX's subsequent use, sublicensing or distribution of the de-identified dataset stays within that agreement.

Frequently asked questions

What is a retail refund policy conflict dataset?

It is a set of real refund requests where at least two rules pointed to different answers, each packaged with the order lines, the policy version in force, the item flags, the jurisdiction facts, the decision and later outcome evidence. It is used to train or evaluate agents that must decide which refund rule controls.

Does the store's return window always control?

No. A statutory right in the customer's location can override it, as with the EU's 14-day withdrawal period for distance sales or California's display duty for no-refund policies. Platform settings such as final sale flags and window overrides also change the answer for specific items.

Which policy version should an agent apply when the policy changed after purchase?

A common approach is to apply the version that was in force and disclosed when the customer bought, but that is a company rule to confirm with counsel. The packet should store the version log with effective dates so the agent can be graded on applying the right one.

Can a resolved helpdesk ticket be used as the label?

Not on its own. A solved status records that a conversation ended, not that the correct rule was applied; the label should come from a reviewed decision with a rule citation, confirmed by the refund or return transaction and the absence of a later dispute.

How many conflict types should a packet cover?

Cover every collision your store actually sees, sampled so each type has enough tasks to measure. Routine in-window returns can be included as controls, but they should not dominate the packet.

Sources

  • The EU Consumer Rights Directive gives consumers 14 days to withdraw from a distance contract and requires reimbursement, including delivery costs, within 14 days of notice, with a right to withhold until goods or proof of return arrive. Source
  • California Civil Code section 1723 requires display of a policy denying full refunds or equal exchanges for at least seven days after purchase and makes a non-compliant retailer liable for the purchase price on returns within 30 days. Source
  • The FTC's business guide says refunds for cancelled mail, internet or telephone orders are generally due within seven working days, or one billing cycle when the seller is the creditor. Source
  • Shopify's return rules apply the shortest window when several overrides cover an item and keep final sale items non-returnable even when an override applies. Source
  • The tau-bench retail policy allows returns only for delivered orders and requires refunds to go to the original payment method or an existing gift card. Source
  • The tau-bench environment rewards a run only when the database state after the agent's actions matches the state produced by the correct actions. Source
  • The tau-bench repository reports agent results as pass^k over repeated trials and notes that its tasks are superseded by the tau2-bench repository. Source

See if your company qualifies

A short company assessment. No data uploads are needed.

See if your company qualifies